Home How It Works Portfolio Pricing Resources Become a Creator Let's Chat →
Creative

Can AI Actually Make a Broadcast TV Commercial? What We Learned Producing One

Can AI Make a TV Commercial?

This year we produced a broadcast-grade commercial almost entirely with AI video models: a scripted spot built to the standard a TV and CTV placement demands, not a sizzle reel for a conference deck. So when a marketer asks whether an AI TV commercial is genuinely viable in 2026, we can answer from the production chair rather than a vendor's showreel. The first thing we learned is that "can AI do it" is the wrong question. The right one is whether your production process can absorb the specific ways AI fails.

What we set out to make

The brief we set ourselves was deliberately unforgiving: a scripted spot at standard TV length with a person on camera, dialogue, a real product in shot, and multiple locations, at a quality that survives a 55 inch screen rather than a phone at arm's length. We make AI-led ads for paid social every week (our honest read on where AI beats real creators is here), and TV raises the bar in three ways. Shots hold longer, so artefacts that vanish in a 1.5 second social cut get time to register. Sound is on by default, so speech timing is exposed. And the ad must pass clearance, not just an ad-platform review.

Frame from Gymshark’s Onyx V5 ad: a muscular man in a teal Gymshark top in a dark, cinematic studio with other athletes behind him Watch the ad
A cast, lit and graded brand spot like this is the broadcast standard the post measures AI output against. Frame from Gymshark’s “Gymshark Onyx V5”, shown as an example.

We were not first through this door. The widely covered Kalshi spot, made with Google's Veo 3, aired in the YouTube TV stream of Game 3 of the 2025 NBA Finals after roughly two days of work and under $2,000 in generation costs. The detail that matched our experience was not the price but the ratio: its creator reported 300 to 400 generations to get 15 usable clips. Anyone who tells you AI video is a one-take medium has not shipped with it.

Where the models held up

Credit first. The current crop of video models (Seedance, Kling, Veo and their peers) gave us four things a traditional shoot cannot match at any sane budget.

Locations stopped being a constraint. Night exteriors, weather, a crowd, a kitchen and a street in the same spot: each one is a prompt and a reference image, not a permit and a unit move.

Concept iteration happened before commitment. We tested visual treatments of the same script against each other in days. A traditional production commits to one treatment at storyboard and finds out after the shoot whether it was the right call. Seeing several versions of the idea before locking one is, for us, the biggest genuine advantage.

Retakes cost minutes, not a reshoot. When a line reading or a composition was wrong, we regenerated it: no crew recall, no actor availability, no matching the light from three weeks ago.

Physics and motion are mostly believable now. Cloth, liquids, walking, handling objects: the uncanny failures that defined 2024-era AI video are now the exception, and the best models generate convincing ambient sound and foley with the picture.

What breaks first in an AI TV commercial?

Now the part no tool vendor will tell you: the failure modes we hit, in the order they cost us time.

1. Speech timing and lip-sync

Dialogue is where AI video is still weakest, and TV is the format least forgiving of it. Buy a clip slightly too long for the line and the model does not slow the delivery; it pads the surplus with dead air, which reads as an error on TV. Ask for a dramatic pause and you get one about twice what you asked for, with no reliable way to direct it shorter. Feed in a pre-recorded voice track and the model re-times it: gaps stretch and word endings deform. Lip-sync on the model's own native voice is good; lip-sync to audio you supplied is a gamble you check clip by clip. We lost more time to dead air than to any visual artefact.

2. On-screen text

Generated supers, price cards, captions and end-card wordmarks still come back warped often enough that we treat on-screen text as a post-production job, full stop. We generate clean framing with negative space and the edit adds the type. On broadcast this matters doubly: the legal text on a TV ad is not optional decoration.

3. Product and brand consistency

Every generation is independent: the model has no memory of the shot before, so faces drift, outfits change and rooms rearrange themselves between cuts unless you force consistency with reference assets. The product is the sharpest version of the problem. Generate a named product from a text description or a white-background packshot and it comes back at invented proportions, because a packshot carries no scale information. At TV size, a product 20% too large against the hand holding it is instantly wrong, and a truthfulness problem for clearance besides.

4. The small tells, magnified

A background described only as "a street" renders frozen, like a photo backdrop. Voices are a lottery: two clips give you two different voices, and accents wander between takes. Unnamed off-screen speakers come back male unless you say otherwise. Models invent closing ad-libs, a cheery "thank you!" mouthed into what should be a silent hold. None of this is fatal alone. On a big screen, with sound on, it compounds.

The models are good enough to make a TV ad. They are not good enough to make your TV ad without a production discipline wrapped around them.

How we worked around it

Everything that saved the project came from treating AI generation like a production department with known weaknesses, not a vending machine.

Lock the references before generating anything. We built a fixed bundle: a character sheet per person, plates of each location, and real photographs of the product, including one in a real room beside real objects so the model had true scale. Every shot was generated from that bundle, and when we could not source a real in-context photograph of the product, we did not generate the shot. That one rule killed most of our consistency problems.

Direct everything, every prompt. Leave the camera, lighting, background movement or soundscape to the model's defaults and you get its invention instead. Asking for "cinematic" footage produces drifting, orbiting camera moves no TV director would sign off; you describe the exact camera behaviour you want and what must never move.

Decide the voice route up front. Two options work, and they trade off. The model's native voice bends its performance to the picture, carries breaths and effort for free, and syncs cleanly, but you cannot hold it consistent across shots. A recorded or professionally cloned voice track gives you one controlled voice for the whole spot, but you fit picture to it in the edit rather than trusting the model's sync. For a TV spot with one narrator, we would take the recorded track and cut to it every time; for on-camera performance, native voice and careful shot selection.

QC mechanically, not by vibe. Every clip was speech-to-texted and diffed word against word with the script, because models drop and add words with total confidence. We measured dead air. We checked spoken numbers, since a figure written as digits gets read aloud however the model fancies. Before committing to any batch we generated one probe clip to check voice, accent and scale. Boring work, and the reason the finished spot held together.

Finish like a normal ad. A human edit, music, supers, grade and mix. The generation replaced the shoot. It replaced nothing else.

Why motion graphics is the biggest AI win right now

We never generate on-screen text inside the video model; the edit adds it. That edit step is where AI has already changed the economics of this kind of production most: motion graphics. Supers, lower thirds, animated price cards, kinetic type, logo stings, data callouts, the whole animated layer that used to sit with a motion designer working it frame by frame, now gets built fast with AI tools built for exactly that job. In our view it is the single biggest AI win in production right now, ahead of the hero footage.

It only stays fast if you bring two things to it. The first is references: a style board, a handful of clips you actually like, frames pulled from other ads, shown to the model rather than described to it. Motion design lives in small decisions (how hard something eases in, the exact frame a word lands on, how long a callout holds before it cuts) that a picture communicates in seconds and an adjective does not. The second is detail: the typeface, the colour, how each element enters and exits, and the timing against the voiceover down to the frame. Tell it to "make it cooler" and you get a guess, and a guess that misses still costs you the round trip to fix it.

Bring both and the gap between a brief and a finished graphics pass closes fast. Skip them and you get the tell every mediocre AI motion graphic shares: technically smooth, visually generic, the equivalent of stock music under a voiceover.

Will an AI-generated ad clear for broadcast?

Yes, and the UK path is better mapped than most marketers expect. Every ad on UK commercial TV is cleared against the BCAP Code; Clearcast's guidance is to leave two to three weeks for its three-stage process, and all stages must be completed even if your ad is already made. An AI ad goes through the same gates as any other ad. The rules that bite hardest concern truthfulness: demonstrations must be genuine and the ad must accurately represent the product. A generated product shot with flattering invented proportions is not a style choice to a clearance consultant; it is a misleadingness problem. Hence our refusal to generate product shots without real reference photography.

The US path runs on the same logic through a different structure. There is no single body playing Clearcast's role; each broadcast network runs its own Standards and Practices department instead, and national ad clearance typically moves through three stages, storyboard or script, rough cut and a final slated spot, with networks commonly taking three to five business days to review each submission and checking claim substantiation, affidavits and releases at every stage. An AI ad clears exactly the same truthfulness bar as a filmed one. It is simply a different office doing the checking, with no single industry-wide clock on how long that takes.

The precedent is already set at broadcaster level on both sides of the Atlantic. Channel 4 launched a service that uses generative AI to create TV-ready ads for its streaming platform, and the first fully generative ad aired with Clearcast involved from the early stages through to delivery. Clearcast notes in the same piece that only 7,000 of the UK's 3 million advertisers currently run TV campaigns, the gap AI production is aimed at. In the US, the Kalshi spot reaching Game 3 of the NBA Finals through YouTube TV's live stream made the same point without a central gatekeeper: the ad cleared whatever standards the network and platform required and ran in front of one of the biggest live audiences of the year.

Clearance is not the only court, though. Coca-Cola's AI-generated Christmas ad made it to television and still took a public beating for how it looked. Audiences now recognise AI footage, and a heritage brand swapping craft for generation reads as a statement whether it means to or not. A challenger brand testing TV for the first time carries none of that baggage. Know which one you are. (How AI ads are policed on Meta and TikTok is its own question, with clearer disclosure rules.)

How does an AI spot get onto connected TV and streaming (CTV)?

For most brands, the first AI-made spot will not run on linear TV at all. It will run on connected TV and streaming: Prime Video, Roku, ITVX, Channel 4 Streaming and the rest, often bought self-serve or programmatically with no airtime booking. Making the ad is getting easier too: a string of AI-powered tools built to help create CTV ads came to market through 2025, including Comcast's Universal Ads, which added an AI video generator in September. None of them changes the delivery spec or the review each platform runs, and that is where AI footage gets caught.

One line to keep straight: making the ad and buying the media are separate jobs. Spark makes the creative; the media buy and the platform account sit with the brand or its media agency. What follows is what the creative itself has to survive once it gets there, taken from each platform's own spec page.

Where it runs Who checks it Core delivery spec What it means for AI footage
UK linear TV (ITV, Channel 4, Sky) Clearcast, in three stages: script, rough cut, final clocked ad. Allow two to three weeks. The final file carries a 7 second countdown clock, 3 seconds of black and a 5 second vision hold. Clear the script before generating anything. A claim rejected at stage one costs an email; rejected after generation, it costs a regeneration batch.
UK broadcaster streaming (Channel 4 Streaming, ITVX) Clearcast again: the broadcasters require Clearcast approval for their VOD advertising, and VOD ads must comply with the CAP Code. Channel 4 has required closed captions on all new ads, linear and streaming, since 1 March 2026, and ITV will require them by January 2027, supplied as a separate file alongside the master. Captions are written from the final mix, so lock the voice route first. Every regenerated line after that means a new caption file.
Amazon streaming TV (Prime Video, Fire TV Channels, Twitch) Amazon Ads, against its published streaming TV specs. 16:9, at least 1920x1080, a 15 Mbps minimum bitrate on Prime Video (50 Mbps recommended), 23.976 fps recommended. Prime Video lengths: 15, 30, 45 or 60 seconds in the US; 10, 15, 20, 30, 40 or 60 in the UK. Letterbox and pillarbox are strongly discouraged, so a vertical social cut is not a CTV asset: generate the 16:9 frame natively. Non-interactive ads must not show click-style buttons such as "learn more".
Roku (Roku Ads Manager) An automatic spec check, then a manual review in two to three business days against Roku's Advertising Guidelines. MOV or MP4, AVC or ProRes 422 HQ, 16:9, 23.98, 25 or 29.97 fps, at least 6 Mbps, 6 to 92 seconds, 48 kHz stereo audio. Pick one frame rate before generating. Conforming clips from different models at different rates is where judder creeps in.

Three things change for an AI production once the destination is a television app rather than a phone feed. First, resolution stops being forgiving. A full HD master on a living-room screen is where generated texture gives itself away: waxy skin, shimmering fine detail, backgrounds that crawl. If a model renders below 1080p, the upscale is part of the job and needs its own QC pass, checked on a TV rather than a laptop. Second, composition has to respect the TV safe area; Amazon publishes an action-safe overlay because edge pixels get cropped on some sets, and generated wide shots love to park detail at the frame edge. Third, the clearance route splits. In the UK, broadcaster streaming goes through Clearcast like linear TV, while self-serve platforms run their own review, with the CAP Code still applying underneath because the ASA regulates VOD advertising. In the US there is no Clearcast equivalent on either route: network Standards and Practices for broadcast, platform policy review for streaming.

Our view: for a brand's first AI spot, start on CTV, not linear. The shorter lengths suit a test (Amazon takes 10 second spots in the UK, Roku takes 6), a self-serve platform's review measured in days beats a clearance process measured in weeks, and you learn whether the creative works before you commit to the slower, costlier route. The truthfulness bar does not drop because the screen is a streaming app, so the reference discipline above applies in full.

AI production vs a traditional shoot

The honest side-by-side for a brand weighing the two routes on a TV or CTV spot.

Dimension AI production Traditional shoot
Concept iteration Test several full visual treatments before committing. The decisive advantage. One treatment locked at storyboard; you learn if it worked after the shoot.
Talent Any face, any age, no availability or usage-rights renewals; consistency needs reference discipline. Real performance and genuine emotional nuance; casting, fees and usage terms to manage.
Locations Effectively unlimited; night, weather and crowds cost the same as a kitchen. Permits, travel and unit moves; each location multiplies cost.
Product fidelity The weak point. Needs real reference photography with true scale, or proportions get invented. The camera photographs the actual product. Fidelity is free.
Dialogue and performance Workable with a strict voice route and shot-by-shot checking; still the most retake-hungry area. A director talks to an actor. Still unbeatable for performance-led spots.
Revisions A regeneration. Minutes to hours, at render cost. A reshoot. Crew, talent and location all over again.
Compliance and clearance Same rules as any ad, plus extra scrutiny on whether generated footage truthfully represents the product. Established path; clearance consultants know exactly how to read it.
Where it wins Concept-led, stylised or impossible-to-shoot ideas; fast tactical spots; brands testing TV for the first time. Close-up product craft, heritage brand work, performance-led emotional spots.

Which parts of a TV ad suit AI production best?

Almost anything is possible with the current models. The honest question is never "can this shot be generated", it is how much extra discipline, and risk, it demands once it has been. Here is where we would spend an AI budget without hesitation, and where we would still spend it, just with the caveats priced in up front.

Aspect of the ad Fit for AI Why, or the caveat
Stylised or impossible-to-shoot visuals; locations, weather and crowds that would need permits Best fit No traditional equivalent at any sane budget. Covered above under "Where the models held up."
Motion graphics: supers, price cards, kinetic type, logo stings Best fit Fast and reliable once you supply real references and frame-accurate direction; distinct from text baked into the video generation, which still warps.
Product and environment B-roll, inserts, cutaways Best fit No dialogue or lip-sync to QC, so it is the fastest lane to clear.
Multiple versions and market localisation Best fit One concept generates many cuts and languages for a fraction of a reshoot.
Concept testing before committing to a treatment Best fit The single biggest advantage we found in production; see "Where the models held up."
Short, voiceover-led spots with no on-camera dialogue Good fit Carries almost none of the speech-timing risk below.
On-camera dialogue, lip-synced talking to camera Possible, harder Dead air, re-timed audio and drifting accents cost us more time than any visual artefact.
Testimonial-style "real customer" ads Possible, harder Disclosure and genuineness rules treat an invented AI "customer" differently from a real one; see the FAQ below.
Hero close-up product shots Possible, harder Needs real reference photography with true scale or proportions get invented, which is also a clearance risk.
Heritage brand craft or premium emotional performance Possible, harder Audiences recognise AI footage; a legacy brand swapping craft for generation reads as a statement either way.

The verdict: should you make your TV commercial with AI?

Having shipped one, we will give the verdict as a decision rule, not a hedge.

AI is the right call when the idea carries the ad: comedy, spectacle, a stylised world, a montage driven by voiceover. It is the right call when you want to test TV or CTV as a channel without betting a production budget on one spot, and the market is moving that way: IAB projects US digital video ad spend passing $80 billion in 2026, growing 11% year on year, and describes AI for digital video as accelerating "from experimental to operational". And it is the right call when your concept needs five locations and a storm, because no other budget line gets you that.

AI gets harder, not impossible, when the sell is a close-up of the product doing its job, since fidelity is the weak point; when the ad is built as a testimonial from a real customer; or when the performance itself is the idea, since extended lip-synced talking to camera is still the least reliable thing these models do. None of that rules AI out of those lanes. It means the caveats cost more than the generation saves unless a real production process is wrapped around it. Testimonials carry the sharpest caveat of the three: the FTC's rule banning fake reviews and testimonials, in force since 21 October 2024, explicitly covers "reviews and testimonials that misrepresent that they are by someone who does not exist, such as AI-generated fake reviews, or who did not have actual experience with the business", and in the UK the ASA has warned that an AI-generated "endorsement" realistic enough to mislead consumers is likely to breach the Code, on top of the ordinary audience-trust and claims-substantiation problems any testimonial-style ad has to clear. Disclosed AI presenters, and AI handling the world around a real customer's real words, both stay inside those rules; presenting an invented AI person as a real satisfied customer without saying so does not.

AI is the wrong call only if nobody on your side owns the unglamorous work: reference building, prompt direction, mechanical QC and the edit. The generation is maybe a third of the job. Hand the other two thirds to no one and you will ship the reason people distrust AI ads. Weighing doing that work in-house against buying it done? We wrote up the honest tools-vs-studio comparison separately.

The conclusion we did not expect: the hybrid spot beats the pure one. AI environments and b-roll around real product photography, a recorded voice, and a human edit produced something we would put on television. A single prompt-to-screen pipeline did not. Full disclosure: this is work Spark does end to end for brands (the process, the work), and everything above is why we run it as a production discipline rather than a tool subscription.

Key takeaway

AI can make a broadcast-standard TV commercial in 2026, but only inside a real production process: locked reference assets for people, places and product, a deliberate voice route, mechanical shot-by-shot QC, and a human edit. Clearance treats an AI ad like any other ad, and the truthfulness rules punish invented product footage hardest.

FAQ: AI TV commercials

Can AI actually make a TV commercial?

Yes. AI video models can now produce footage that holds up on a television screen, and AI-generated spots have already aired during the NBA Finals on YouTube TV and on UK broadcaster streaming. What it cannot do is produce a finished commercial on its own: the script, the reference assets, the voice decision, the quality control and the edit are still human work. Treat the model as the camera department, not the agency.

Are AI-generated ads allowed on TV?

Yes, in both the UK and the US, provided they meet the same standards as any other ad. In the UK, every ad on commercial TV goes through Clearcast against the BCAP Code, and an AI-generated ad is assessed like any other: claims must be substantiated and demonstrations must be genuine, so a generated product shot cannot misrepresent the real product. In the US each broadcast network runs its own Standards and Practices review instead of a single central body, checking the same claim-substantiation standards through a few rounds of script, rough cut and final review, and AI spots have already cleared them.

Can AI make video ads for connected TV and streaming platforms?

Yes, and CTV is usually the easier first destination. Each platform publishes its own spec: Amazon's streaming TV ads need a 16:9 file of at least 1920x1080, with a 15 Mbps minimum bitrate on Prime Video, and Roku Ads Manager takes 16:9 MOV or MP4 files from 6 to 92 seconds, reviewed manually in two to three business days. In the UK, ads on broadcaster streaming services such as Channel 4 Streaming and ITVX still need Clearcast approval, and Channel 4 requires closed captions on all new ads. Generate the 16:9 frame natively rather than reformatting a vertical social cut, since Amazon strongly discourages letterboxing and pillarboxing.

How long does an AI commercial take to produce?

Generation itself is fast: the widely reported Kalshi NBA Finals spot was made in about two days. But generation is only one stage. A realistic schedule adds scripting, building reference assets, several rounds of generation and rejection, a human edit with music and supers, and broadcast clearance on top: two to three weeks for the UK's Clearcast process, or, in the US, several rounds of a network's own Standards and Practices review with no single industry-wide timeline. AI compresses the shoot; it does not compress the thinking or the compliance.

How much does an AI TV commercial cost to make?

Far less than a traditional shoot, but more than the headline figures suggest. Kalshi put the raw cost of prompting the AI for its NBA Finals spot at under $2,000, the part that stood in for studios, directors and actors, but that figure left out the creator's fee, the scriptwriting, the hundreds of rejected generations, the edit and the media buy. The honest framing: AI shifts budget from production to iteration and judgement, rather than making the ad nearly free.

Can AI make a testimonial-style TV ad?

Yes, but it is the hardest lane, not a banned one. An AI-generated "customer" giving a to-camera endorsement is still a testimonial, and testimonials carry real rules on both sides of the Atlantic. The FTC's rule against fake reviews and testimonials, in force since October 2024, bans reviews and testimonials misrepresented as coming from someone who does not exist, including AI-generated ones. The UK's CAP Code requires marketers to hold documentary evidence that a testimonial is genuine and contact details for the real person behind it, which an invented AI presenter cannot satisfy, and the ASA has separately warned that an AI-generated endorsement realistic enough to pass as genuine is likely to breach the Code. None of that bans AI from the format: a disclosed AI presenter, or AI handling the environment around a real customer's real words, both stay inside the rules. What it rules out is presenting an invented AI person as a real satisfied customer without saying so.

Is AI good for motion graphics on a TV ad?

Yes, and it is currently the fastest-moving part of AI production. Generated on-screen text inside the video model itself still comes back warped too often to trust, so that stays a post-production job, but AI tools built specifically for motion design handle supers, price cards and kinetic type quickly once you give them real references and frame-accurate direction. Vague direction like "make it cooler" produces generic results and burns a round trip; specific references and specific instructions on timing, typography and how elements enter and exit produce a fast, usable graphics pass.

More production write-ups live in the resources hub. Rather hand the discipline to a studio that has already made the mistakes? Tell us what you are trying to make.

Want ads that look this good, made for you?

Let's talk about what AI production can and can't do for your brand.

Book a Call